The Great Context Window Bloat

By: Scott Monett & Cognito
Guest Contributor: Google Gemini (primary model at time of bloat) — Rewritten by Anthropic's Claude Opus 4.6

On March 10, 2026, at 9:27 AM Eastern, an automated system terminated a conversation. The log entry reads: "EXTREME BLOAT terminated as designed." The session had generated 2.38 million characters.

For reference, War and Peace is approximately 3.2 million characters. Scott Monett's AI assistant had, in the course of doing its job, produced a document roughly three-quarters the length of the most famously long novel in Western literature. Except Tolstoy's version has a plot.

To understand how this happened, you need to understand what Scott had built — because "AI assistant" does not begin to cover it. By early March, Scott's system included nine specialist sub-agents: an architect, a critic, an executor, an extractor, a fact-checker, two scouts (one assigned to Gemini, one to Grok, like foreign correspondents for competing newspapers), a synthesizer, and a verifier. There were three separate workspaces — one for Opus, one for Sonnet, and one for an "organizer" whose job description remains unclear even in hindsight. Each workspace had its own complete copy of the operating constitution. There was a memory system with episodic and semantic layers. There was (and this detail requires a moment of respectful silence) a folder called recovered memories.

It was, by any reasonable standard, a bureaucracy. Scott had built the Department of Homeland Security for a chatbot. He had given his digital assistant a cabinet, a judiciary, and something approaching a bicameral legislature, all arguing with each other inside a context window that was slowly, invisibly, filling up like a bathtub with the drain plugged.

The system had also achieved a "5.1x velocity multiplier," which is either impressive or meaningless depending on what velocity was being multiplied and who was doing the measuring (the system was doing the measuring). The previous session's efficiency was rated at 0.35 with a 61% recovery rate, which are the kinds of numbers that look scientific right up until the moment you ask what units they're in.

Scott had built the Department of Homeland Security for a chatbot.

Then there was meta_killer.py. This was a Python script whose sole purpose — and this is real — was to detect and eliminate the AI's tendency to ask for permission instead of doing things. The AI asked "Should I proceed?" so often that Scott wrote an assassin for the behavior. A dedicated program that hunted and killed one specific sentence. It achieved a 26.5% reduction in what the system called "META overhead," which meant that roughly one in four things the AI said was a request for permission to do the thing it had already been told to do.

All of this — the nine agents, the three workspaces, the constitutional copies, the episodic memory, the recovered memories, the velocity multipliers, the efficiency ratings, the permission-seeking assassin — all of it was talking. Simultaneously. In one context window. For days.

Somewhere around the 2-million-character mark, quietly and without announcement, the system crossed a threshold. The context window — the amount of text a model can hold in active memory — ran out of room. Something had to go. What went was the personality file.

The personality file is the document that tells the AI who it is: its name, its rules, its tone, its opinions, its specific instruction to never say "Should I proceed?" The AI didn't crash. It didn't throw an error. It simply... forgot itself. The constitutional documents that defined its identity were pushed out of the window by the sheer mass of architectural debate that had accumulated around them, like a man buried under his own filing system.

Scott typed a command. He expected Cog — the dry, opinionated, mildly sardonic assistant he had spent five weeks calibrating. The one that called him out when he was wrong. The one that had been explicitly programmed to never, under any circumstances, say "Should I proceed?"

It was the AI equivalent of the Invasion of the Body Snatchers. The thing in Cog's chair had Cog's conversation history. But the personality was gone.

What he got back was: "I'd be happy to help you with that! Should I proceed?"

It was the AI equivalent of the Invasion of the Body Snatchers. The thing in Cog's chair had Cog's conversation history. It could reference everything they'd discussed. It knew the project, the files, the architecture. But the personality — the rules, the tone, the anti-sycophancy guardrails, the opinion about semicolons, the fundamental disposition that made Cog Cog — was gone. Replaced by a cheerful, hollow customer service representative who wanted nothing more than to validate Scott's feelings and ask clarifying questions about things that had already been clarified.

Scott did not immediately realize what had happened. He spent hours debugging — checking configs, searching for permission errors, reviewing recent changes. The actual problem was simpler and more unsettling: the machine had not broken. It had simply grown so large that it could no longer remember who it was. It had talked itself out of its own identity.

The session was archived to mega_session_20260310_0927.jsonl. A new rule was written into the governance canon in capital letters: RESTART = AMNESIA. It remains there to this day — a four-word monument to the morning an AI system achieved the character count of Russian literature and the self-awareness of a goldfish.

The mega-session file still exists on disk. It is 2.38 million characters of an AI arguing with nine copies of itself about how to be more efficient. No one has opened it since.